LLM agents fulfill natural‑language requests by invoking tools, data sources, and other agents, and even simple requests often require several interdependent calls. Conventional approaches let the LLM select calls by trial and error, offering no guarantee of an executable plan. This paper introduces Capability‑Grounded Agentic Planning (CGAP), which first compiles the tool catalog deterministically into a PDDL model, then uses an LLM to translate the request into goals, and finally solves the model with a classical planner. Because tool outputs are unknown beforehand, the model abstracts only the availability of inputs, not their concrete values. During execution, an LLM adds entity‑specific calls omitted by the abstract plan, and deterministic checks reject any call that would break the plan. Every argument is drawn from a cited source such as a conversation span or a prior tool response, and a result is considered delivered only after an LLM verifier confirms it matches the request. Experiments on ToolSandbox, ComplexFuncBench, BFCL multi‑turn, and $\tau^2$-bench retail benchmarks show CGAP outperforms a standard tool‑calling agent by up to 28.9 points.
Review: CGAP’s formulation of tool usage as a planning problem yields more reliable multi‑step reasoning and markedly improves task success rates.